Adaptive Physics Refinement at the Microstructure Scale

Abstract

A long-sought goal of computational materials science and engineering has been a simulation framework that spans all necessary length and time scales, potentially from electronic structure to structural engineering, providing the appropriate level of physics fidelity where needed and enabling the user to trade off accuracy and computational time in an optimal manner. Analogous to adaptive mesh refinement methods that dynamically (and automatically) coarsen and refine a computational mesh based on local requirements, adaptive physics refinement methods utilize higher-fidelity physics models as needed, e.g., replacing a phenomenological constitutive model with a direct polycrystal plasticity simulation. While there have been several demonstrations of similar concurrent multiscale methods over the past 20–30 years, only now is a more general capability becoming viable due to advances in algorithms and computer architectures and middleware. In this chapter, we briefly review this history, focusing on two methods in particular: the heterogeneous multiscale method and adaptive sampling. The computational workflow, data, and runtime requirements of these methods are used to identify key enabling technologies that have recently gained widespread adoption, including task-based programming models, heterogeneous computer architectures, database, and machine learning algorithms.